AION On-PremiseThe question stays in-house. So does the answer.

AION On-Premise puts optimized models, Agents and your integrations on a machine that's yours. Follow a request from start to finish: the model that answers always stays inside the company boundary, and the internet only comes into play when needed.

Request a demo
Agents · Purchasing profilein-house

How much notice do we need to give to terminate the plant maintenance contract?

You need to give 90 days' written notice, sent by certified email (PEC) to the supplier's contracts office (art. 14.2).

  • MCP · contract archive
  • embeddings + reranker
  • source: Contratto_manutenzione.pdf · p. 7

01 · Agents

01

The question reaches Agents.

Chat, agents, memory and MCP tools, with profiles and users for each department, a REST API and a console to manage it all. It's open source: you don't pay for a separate software license.

  • If you prefer, it can run on a separate application server, on your local network
Agents in a conversation: an answer with a table, the cited web sources and the generated Word report, with the session files on the right
Agents at work: a search with cited web sources and the generated Word report.
02

It searches your systems, and the web when needed.

With MCP connectors, the agent consults the systems you choose to connect. If you ask, or when it helps, it also searches the internet with Google: the reasoning, though, always stays with the model on your PGX. PGX installations also include n8n, with the AION node, if you need automations.

  • Document archiveMCP
  • DatabasesMCP
  • Management software and ERPMCP
  • TicketingMCP
  • Logs and monitoringMCP
Internet search (Google)on request

n8n with AION node

We develop MCP integrations with your systems together, around your project.

03

The answer comes from a model running on your machine.

AION Hub is the catalog of models on your machine: over 30 open-source models that AION has quantized, adapted and optimized. Start and stop on the GPU only the ones you need. Behind a single OpenAI-compatible gateway, the language model, embeddings, reranker and OCR work together.

An NVIDIA DGX Spark system (Lenovo ThinkStation PGX), the machine AION On-Premise runs on
NVIDIA DGX Spark · GB10 Grace Blackwell · 128 GB of unified memory · pictured: Lenovo ThinkStation PGX
Platform
NVIDIA DGX Spark
AION Hub catalog
Over 30 open-source models
External providers called
None

Models are updated via VPN. If the machine is isolated from the network (air-gapped), a USB stick is all it takes.

Same path, bigger machine.

Hub, models and Agents stay the same. As people and questions grow, the only thing that changes is the machine.

Choose a configuration to see its details

AION On-Premise Level 1

PGX

Level 1
PGX machines
1
active users (estimate)
~10
concurrent requests
8

The starting point for teams and departments.

PGX: systems built on NVIDIA DGX Spark, such as Lenovo ThinkStation PGX. NVIDIA GB10 Grace Blackwell chip, 128 GB of unified memory. Installation included, one license per machine. What the machine looks like inside

Open-source models, optimized by AION

AION does not train models from scratch: it starts from the best open-source models of the leading providers and gets them ready to run well on your hardware: quantizing, adapting and optimizing them.

  1. Open-source base

    We start from the leading providers' models: a solid, transparent foundation.

    • Qwen
    • GLM
    • Kimi
    • Mistral
  2. AION quantization

    The model is 69–70% lighter, with an estimated quality loss between 1% and 3%.

  3. Fine-tuning

    Where it helps, we adapt it to your field, to the right tone and to the tasks it has to handle.

  4. In the AION Hub catalog

    Over 30 models ready for NVIDIA DGX Spark: you start them from AION Hub and they integrate with Agents.

No usage limits
Query the models as much as you need, with no caps.
Always up to date
As long as you have the AION On-Premise license, the catalog models stay up to date.

What your company knows doesn't go into the models: documents are consulted with cited sources, so no sensitive data ends up in the weights.

Pipeline, quantization and benchmarks of AION models
04

The answer comes back with its sources.

Before answering, the agent looks into the systems you've connected: document archive, database, company mail, network folders. Whoever asked gets an answer that cites every source consulted. If you ask, or when needed, it also searches the internet.

What the agent consulted

Company boundary

Question: How much notice do we need to give to terminate the plant maintenance contract?

Agents

Purchasing profile

Queries the connected systems

  1. Document archive

    MCP

    Maintenance_contract.pdf · p. 7

    90 days' written notice, by certified email (art. 14.2).

  2. Management database

    SQL

    contracts table

    Expires December 31 · automatic renewal active.

  3. Company mail

    MCP

    Purchasing mailbox

    Supplier email of March 3: renewal terms unchanged.

  4. Company files

    network

    \\fileserver\Purchasing\Templates

    Termination_letter_contracts.docx · termination letter template.

Answer

You need 90 days' written notice, by certified email . The contract renews on December 31 : the letter must go out by October 2. The supplier has already confirmed the renewal terms  and the letter template is in the Purchasing folder .

Internet searchon requestNot needed for this question.

The same models can also serve your applications: any OpenAI-compatible tool connects to the gateway.

  • n8n
  • Open WebUI
  • LangChain
  • LlamaIndex
  • Flowise
  • AnythingLLM

Third-party tools that work with OpenAI-compatible APIs. n8n is already installed on the PGX.

05

Need an external model? The Router decides.

For models, the path reaches the boundary only when you decide it should. AION Router puts local and remote models behind a single OpenAI-compatible endpoint and, on every request, decides which model answers, which data can leave and how much it costs.

Privacy
Masks personal data before it reaches a remote provider.
Costs
Picks the cheapest model that can answer, with a budget for each team.
Control
API keys that open only the approved models, routing rules and usage reports.
How the Router decides, switch by switch

One PGX is enough to start. The others arrive when you need them.

You don't have to size tomorrow's infrastructure today. With just one machine the request path is already complete: Hub, AION models and Agents. If usage grows, you add another next to the first.

  1. PGX

    Start hereLevel 1

    ~10 active users · 8 concurrent requests

    When it makes sense

    For a team or department starting to use the agent on its own documents and systems. Everything's already there, on a single machine.

    What changes

    Nothing to add: Agents, models and integrations all run on this machine, inside your network.

  2. 2 × PGX

    Level 2

    ~30 active users · 24 concurrent requests

    When it makes sense

    When the agent moves beyond one department and the whole company uses it, and at peak hours requests start to pile up.

    What changes

    You put a second machine next to the first, connected over a high-speed network. The request path doesn't change.

  3. 3–4 × PGX

    Per project

    When it makes sense

    When several departments work together, or you need bigger models and heavier workloads.

    What changes

    The model catalog includes three- and four-machine configurations. They are not an automatic step: we size them with you around your project.

  4. Workstation and Server

    On quotation

    When it makes sense

    When the workload needs dedicated NVIDIA GPUs, or the AI has to be integrated into your datacenter.

    What changes

    These are project choices, not the next step after the PGX. We size them around your case, and only if you really need them.

Tell us how many people will use the agent and for what: together we'll work out whether one PGX is enough for you.

Users and requests are estimates: real usage depends on how you work. Installation included, one license per machine.

Getting started. What's included is up to you.

  1. Included

    Installation

    We install and configure the platform on your machine.

  2. Optional

    Training

    The Professional AI course (16 hours) for those who use the agent, the Admin AI course (8 hours) for those who manage it.

  3. Optional

    Support

    A package of support hours after launch.

  4. Per project

    MCP integrations

    Connectors to your systems, built around your use case.

Let's talk about your project.

We'll show you AION On-Premise on a real use case and size the right machine together.

Headquarters
Corso Vercelli 89, 28100 Novara

Your data is treated according to our Privacy Policy and will not be shared with third parties.